I've been getting some of my own concoctions running on AEMR (well, at least, hitting some problems further down the line). Here's how I proceed.
The input to AEMR is one .jar file with Job, Mapper, Reducer, support code, etc., yes? So everything has to go in there, for sure. Do I understand correctly that there are multiple .jars in play here? packaged up into one big .jar, trying to use the MANIFEST.MF classpath mechanism? Can you do that -- I did not think this mechanism could refer to stuff inside its own self. I just went straight to repacking everything into one .jar with an Ant script. There should not be multiple classloaders in use by Hadoop. Or if somehow it does, well your classes are in one .jar and so should all end up in the same classloader no matter what. So I suspect something along the above lines.
